[KimiViT][Perf] Fuse per-layer QK RoPE into one in-place kernel (#58651)

This commit is contained in:
Thien Tran
2026-09-27 18:18:06 -07:00
committed by GitHub
parent 3184226984
commit 39d49539de
3 changed files with 140 additions and 7 deletions
@@ -225,3 +225,46 @@ def test_apply_rotary_emb_xpu_matches_native(
expected = op.forward_native(x, cos, sin)
actual = op.forward_xpu(x, cos, sin)
torch.testing.assert_close(actual, expected, atol=0, rtol=0)
@pytest.mark.skipif(not current_platform.is_cuda_alike(), reason="CUDA/ROCm only test.")
@pytest.mark.parametrize("num_tokens", [257, 2304])
@pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16])
def test_packed_qk_rope_correctness(
num_tokens: int, dtype: torch.dtype, default_vllm_config
):
"""packed_qk_rope_ must be bitwise identical to the per-tensor
ApplyRotaryEmb path with enable_fp32_compute=True, and must leave the V
slice untouched."""
from vllm.model_executor.layers.rotary_embedding.common import ApplyRotaryEmb
from vllm.model_executor.layers.rotary_embedding.packed_qk_rope import (
packed_qk_rope_,
)
num_heads, head_dim = 12, 128
rng = torch.Generator(device="cuda").manual_seed(0)
angles = (
torch.rand(num_tokens, head_dim // 2, device="cuda", generator=rng) * torch.pi
)
freqs_cis = torch.polar(torch.ones_like(angles), angles)
cos = freqs_cis.real.contiguous()
sin = freqs_cis.imag.contiguous()
xqkv = torch.randn(
num_tokens, 3, num_heads, head_dim, dtype=dtype, device="cuda", generator=rng
)
xqkv_ref = xqkv.clone()
xq, xk, xv = torch.unbind(xqkv_ref, dim=-3)
op = ApplyRotaryEmb(
enforce_enable=True, is_neox_style=False, enable_fp32_compute=True
)
# op() dispatches to forward_cuda / forward_hip per platform.
xq_ref = op(xq, cos, sin)
xk_ref = op(xk, cos, sin)
packed_qk_rope_(xqkv, freqs_cis)
torch.testing.assert_close(xqkv[:, 0], xq_ref, atol=0, rtol=0)
torch.testing.assert_close(xqkv[:, 1], xk_ref, atol=0, rtol=0)
torch.testing.assert_close(xqkv[:, 2], xv, atol=0, rtol=0)
@@ -0,0 +1,81 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""In-place interleaved RoPE on the Q/K slices of a packed QKV buffer."""
import torch
from vllm.triton_utils import tl, triton
# seqlen is not specialized: image sizes vary per request.
@triton.jit(do_not_specialize=["seqlen"])
def _packed_qk_rope_kernel(
qk_ptr, # (seqlen, 2 * nheads, headdim) strided view of the packed QKV
freqs_ptr, # (seqlen, rotary_dim) fp32 re/im view of the complex freqs
seqlen,
qk_row_stride,
qk_head_stride,
rotary_dim: tl.constexpr,
BLOCK_M: tl.constexpr,
):
BLOCK_K: tl.constexpr = triton.next_power_of_2(rotary_dim)
pid_m = tl.program_id(0)
pid_h = tl.program_id(1)
rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
rk = tl.arange(0, BLOCK_K)
mask_m = rm < seqlen
base = qk_ptr + pid_h * qk_head_stride + rm[:, None] * qk_row_stride
mask_x = mask_m[:, None] & (rk < rotary_dim)[None, :]
x = tl.load(base + rk[None, :], mask=mask_x, other=0.0).to(tl.float32)
f = tl.load(freqs_ptr + rm[:, None] * rotary_dim + rk[None, :], mask=mask_x)
cos, sin = tl.split(tl.reshape(f, (BLOCK_M, BLOCK_K // 2, 2)))
x0, x1 = tl.split(tl.reshape(x, (BLOCK_M, BLOCK_K // 2, 2)))
o0 = x0 * cos - x1 * sin
o1 = x1 * cos + x0 * sin
tl.store(base + rk[None, :], tl.interleave(o0, o1), mask=mask_x)
def packed_qk_rope_(xqkv: torch.Tensor, freqs_cis: torch.Tensor) -> None:
"""Rotate Q and K in place inside a packed QKV buffer.
Args:
xqkv: contiguous (seqlen, 3, nheads, headdim); only the Q and K
slices are rotated.
freqs_cis: contiguous (seqlen, headdim // 2) complex64 rotary freqs,
read through its interleaved re/im fp32 view.
Bitwise identical to ``ApplyRotaryEmb(enable_fp32_compute=True)`` applied
to Q and K separately, but in one kernel with ~5x less memory traffic.
Requires triton: callers must check HAS_TRITON and use the unfused path
otherwise. Contract violations raise.
"""
assert xqkv.ndim == 4 and xqkv.size(1) == 3, xqkv.shape
assert xqkv.is_contiguous(), xqkv.shape
assert freqs_cis.ndim == 2 and freqs_cis.is_contiguous(), freqs_cis.shape
assert freqs_cis.dtype == torch.complex64, freqs_cis.dtype
seq_length, _, num_heads, head_dim = xqkv.shape
rotary_dim = freqs_cis.size(-1) * 2
assert freqs_cis.size(0) == seq_length, (freqs_cis.shape, xqkv.shape)
assert rotary_dim == head_dim, (freqs_cis.shape, xqkv.shape) # full rotation
qk = xqkv.as_strided(
(seq_length, 2 * num_heads, head_dim),
(3 * num_heads * head_dim, head_dim, 1),
)
BLOCK_M = 4
# seqlen on grid.x: HIP caps gridDim.y at 65535.
grid = (triton.cdiv(seq_length, BLOCK_M), 2 * num_heads)
_packed_qk_rope_kernel[grid](
qk,
freqs_cis.view(torch.float32),
seq_length,
qk.stride(0),
qk.stride(1),
rotary_dim=rotary_dim,
BLOCK_M=BLOCK_M,
num_warps=2 if rotary_dim <= 64 else 4,
)
+16 -7
View File
@@ -30,6 +30,9 @@ from vllm.model_executor.layers.linear import (
)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding.common import ApplyRotaryEmb
from vllm.model_executor.layers.rotary_embedding.packed_qk_rope import (
packed_qk_rope_,
)
from vllm.model_executor.models.utils import maybe_prefix
from vllm.model_executor.models.vision import (
is_vit_use_data_parallel,
@@ -37,6 +40,7 @@ from vllm.model_executor.models.vision import (
)
from vllm.platforms import current_platform
from vllm.transformers_utils.configs.kimi_k25 import KimiK25VisionConfig
from vllm.triton_utils import HAS_TRITON
from vllm.utils.torch_utils import async_tensor_h2d
logger = init_logger(__name__)
@@ -451,14 +455,19 @@ class MoonViTEncoderLayer(nn.Module):
)
# xqkv: (seqlen, 3, nheads, headdim)
xqkv = xqkv.view(*qkv_shape)
xq, xk, xv = torch.unbind(xqkv, dim=-3)
_apply_rope_input_validation(xq, rope_freqs_cis)
_apply_rope_input_validation(xk, rope_freqs_cis)
rope_cos = rope_freqs_cis.real.contiguous()
rope_sin = rope_freqs_cis.imag.contiguous()
xq = self.apply_rotary_emb(xq, rope_cos, rope_sin)
xk = self.apply_rotary_emb(xk, rope_cos, rope_sin)
# xq/xk alias xqkv, so the in-place rotation below is visible through
# them; only the fallback needs to rebind.
xq, xk, xv = torch.unbind(xqkv, dim=-3)
if HAS_TRITON:
packed_qk_rope_(xqkv, rope_freqs_cis)
else:
_apply_rope_input_validation(xq, rope_freqs_cis)
_apply_rope_input_validation(xk, rope_freqs_cis)
rope_cos = rope_freqs_cis.real.contiguous()
rope_sin = rope_freqs_cis.imag.contiguous()
xq = self.apply_rotary_emb(xq, rope_cos, rope_sin)
xk = self.apply_rotary_emb(xk, rope_cos, rope_sin)
if max_seqlen is None:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()